Kimi K3.1 Pro makes open weights a real production contender
Moonshot's new agentic model is live by API, independently ranks near the frontier, and promises full weights on November 4. The right move is a shadow test, not a rushed migration.
What launched
On October 23, Moonshot released Kimi K3.1 Pro and a Highspeed endpoint across Kimi products and the API. The model accepts text and images, carries a 1-million-token context window, and is due to receive full weights by November 4, 2026.
The independent signal
Artificial Analysis placed K3.1 Pro fourth on its Intelligence Index and reported the highest Terminal-Bench 4.0 result in its evaluation. Its first-party rates are $2 per million input tokens, $10 for output and $0.20 for cache hits, each 33% below K3.
Do not migrate yet
The cheaper rate card did not reduce Artificial Analysis's cost per task: the model produced more output tokens. We would shadow-route real jobs, compare completed-task cost, latency and retries, then revisit migration after the weights arrive and the licence terms can be reviewed.
Kimi K3.1 Pro is the first open-weight contender in this cycle that we would put into a serious production bake-off, but we would not move a live system to it this week. The API is available now and the independent results are strong, especially on agentic terminal work, yet the main strategic benefit — portable weights — is still promised for November 4 rather than in hand. More importantly, lower token prices do not automatically mean lower operating cost: Artificial Analysis found roughly unchanged cost per task because K3.1 Pro generated more output. We would shadow-route representative coding and multi-step agent jobs beside the incumbent, score completed work rather than benchmark points, and log latency, retries, tool failures and total spend. We would only cut over where it wins on that full path, and we would avoid the higher-priced Highspeed endpoint unless reduced waiting time materially improves completion economics. The launch changes the shortlist; it does not remove the need for evidence.
Turn insight into a working system.
We build a shadow-routing and evaluation layer that runs Kimi K3.1 Pro beside your current model, measuring task completion, latency, retries and spend without changing production behaviour. If it passes, we add controlled cutover, fallback, permissions and audit. Talk to us if your team runs coding or agent workflows this release could change.